Rebuilding the Government of Canada's AI assistant
Shared Services Canada had a working AI assistant, but employees tried it once and went back to email. The problem wasn't the model. It was everything around it.
This work was completed outside of SteelSprint.
The Organization
Shared Services Canada (SSC) provides IT infrastructure and services to the Government of Canada. Their Digital Technology Operations branch built an internal AI assistant, the SSC Assistant, to help employees find information across government systems and internal documentation.
The Problem
The SSC Assistant already existed. The AI worked. GPT-4o handled questions well enough. But usage was low. Employees tried it once, found the interface hard to navigate, and went back to email or searching manually.
The issue wasn't the model. It was everything around it: navigation, conversation management, search suggestions, and basic usability. The team needed someone who could rebuild the frontend and ship fast.
The Approach
SteelSprint's founder joined the SSC team as a contractor with security clearance. The focus was on two things: make the interface usable, and connect the assistant to the information employees actually needed.
A complete interface rebuild. New homepage, redesigned menus, conversation history with clear session management, and a full layout overhaul. The goal was to make the assistant feel like a tool people reach for, not a demo they try once.
The assistant was connected to SSC's internal wiki. When employees asked a question, the assistant could surface relevant articles alongside its AI-generated response. This turned the assistant from a chatbot into a search tool that actually knew where things lived inside the organization.
Government software has strict accessibility requirements. Tab order was reworked, keyboard navigation added, and proper ARIA titles applied to every interactive element. This wasn't just compliance. Accessibility improvements made the interface better for everyone.
Technical detailsThe assistant runs on Azure with OpenAI for language and Azure Search for retrieval. SteelSprint's founder rebuilt the frontend, added a wiki search layer, reworked accessibility across the entire interface, and added document upload for context-aware responses.+
TypeScript and Python on Azure. Azure OpenAI for the language model. Azure Search for RAG (retrieval-augmented generation) across government documentation.
A new API endpoint that queries SSC's internal wiki and returns relevant articles based on the user's question. Results surface alongside the AI-generated response, giving employees both a direct answer and the source documentation.
Full tab order rework, keyboard navigation for all interactive elements, ARIA labels and titles across the entire interface. Document upload support for context-aware responses. Prompt engineering refinements for better, more relevant AI outputs.
What was delivered
Over six months, roughly 30 pull requests covering:
Complete interface rebuild: new homepage, navigation, menus, conversation history, profile system
Wiki search: the assistant surfaces relevant internal articles alongside its responses
Document upload: users can upload documents for context-aware responses
Improved AI responses: refined how the assistant interprets and answers questions
Accessibility overhaul: full keyboard navigation and screen reader support across the interface
The Results
Usage increased 50% within three months of the changes going live. The AI model didn't change. The infrastructure didn't change. What changed was that people could actually use the thing.
An internal user group was founded that grew to 100+ members, creating a continuous feedback loop between the people using the assistant and the team building it.
Findings on building autonomous, multi-turn agents were presented to 80+ Architects, Directors, and Senior Directors at the AI Centre of Excellence.
The Takeaway
AI adoption is a design problem, not an AI problem. Most organizations deploying internal AI tools build the model and forget the interface. The SSC Assistant already had a working AI. Usage was still low. An interface rebuild and wiki integration drove a 50% increase without changing the model. If your AI tool isn't getting used, the answer probably isn't a better model. It's a better interface.